Performance Evaluation of Scheduling Mechanisms for Broadband Networks

Transcription

1 Performance Evaluation of Scheduling Mechanisms for Broadband Networks Gayathri Chandrasekaran Master s Thesis Defense The University of Kansas Committee: Dr. David W. Petr (Chair) Dr. Joseph B. Evans Dr. David L. Andrews Information & Telecommunications Technology Center (ITTC), EECS, University of Kansas

2 Abbreviations HF 2 Q - Hybrid FIFO Fair Queuing WFQ - Weighted Fair Queuing FIFO - First In First Out GPS - Generalized Processor Sharing BWA - Broadband Wireless Access DOCSIS - Data Over Cable Service Interface Specification MAC - Medium Access Control QoS - Quality of Service 2

3 Presentation Outline Introduction & Motivation Scheduling Mechanisms & Fair queuing Description of HF 2 Q Performance Characterization of HF 2 Q DOCSIS 1.1 Operation Performance Evaluation of DOCSIS 1.1 Summary of Results Contributions & Future Work 3

4 Introduction & Motivation Future broadband networks will support multiple types of traffic over single physical infrastructure Issue of scheduling mechanisms and bandwidth allocation play a critical role Packet Fair Queuing algorithms have received a lot of attention Hybrid FIFO Fair Queuing (HF 2 Q) a new queuing discipline Has very interesting properties and hence a performance characterization would prove worthwhile 4

5 Introduction & Motivation (cont.) DOCSIS de facto standard for delivering broadband services over HFC networks, defines MAC scheduling mechanisms for BWA, developed by CableLabs The IEEE protocol was developed for Broadband Wireless Access (BWA) systems to provide Internet access and multimedia services to end users. Consolidation of two proposals, one of which was based on DOCSIS 1.1 DOCSIS 1.1 built on top of DOCSIS 1.0, specifies QoS components in terms of MAC scheduling based on underlying traffic requirements Standardized in 2001 a performance evaluation is significant in terms of understanding the operation of the protocol as such and its new QoS features 5

6 Scheduling Mechanisms and Fair Queuing Queuing algorithms determine the way packets from different sources interact Controls order in which packets are sent In the classic FIFO the order of arrival completely determines bandwidth allocation Fair in the sense that all packets get equal mean waiting times Protection against ill-behaved sources is required and hence the requirement of an algorithm that provides fair bandwidth allocation A class of fair queuing algorithms have been proposed since then 6

7 Background on FQ policies FQ algorithms maintain separate queues for packets from individual sources Queues are serviced in a round robin manner Provide fair treatment for supported flows by splitting bandwidth based on pre-defined weights Fairness Criterion : For any two backlogged flows, each flow s service (normalized to its weight) should be nearly the same 7

8 Weighted Fair Queuing WFQ is a packet scheduling technique allowing guaranteed bandwidth services WFQ is a GPS approximation for packet networks GPS is an idealized fluid model that cannot be implemented practically The time at which a packet would complete service in GPS is computed and the packet is assigned a timestamp Timestamps are virtual finish times Packets are served in increasing order of their timestamps 8

9 WFQ : Virtual time WFQ operation linked to the GPS system using virtual time defines the order in which packets are served Virtual time v(t) is a piecewise linear increasing function of time v(t) x x x x x Has a rate of increase inversely proportional to the sum of the service rates of the backlogged flows C v( t2) v( t1) =.( t2 t1) r t k B( t 1, t k ) 2 9

10 WFQ : Virtual Start and Finish Times i th packet of flow k is denoted by p ki, its arrival time by a ki and its length by L k i v(a ki ) is the value of v(t) at the time of packet arrival Virtual finish times are assigned as shown Virtual start time denoted by S i k varies depending on whether a new arrival is to an empty flow or backlogged flow WFQ serves packets in increasing order of finish times 10

11 HF 2 Q description A new and an interesting queuing algorithm discovered while developing the WFQ simulation model A small change in the WFQ operation leads to HF 2 Q Named Hybrid FIFO Fair Queuing since it exhibits desirable properties of both FIFO and WFQ Exhibits FIFO behavior when the total system load is less than unity Behaves as WFQ when the system is overloaded 11

12 HF 2 Q implementation v(t) represented by roundnumber ( currenttime lastrupdatetime) roundnumbe r = roundnumber +. capacity weightsum Instants when the roundnumber (v(t)) is updated is kept track of using a variable lastrupdatetime The lastrupdatetime is not updated when a new packet arrives after an idle period in HF 2 Q Scheduling order thus changes 12

13 HF 2 Q vs. WFQ Virtual Time τ e - end of a server busy period, τ s - instant when the first packet arrives after an idle period from flow a 1, τ 2 - the instant of arrival of second packet Offset between v(t) and v (t) v (t) The new arrival at τ 2 gets assigned v (t) in HF 2 Q rather than v(t) in WFQ Virtual time v(t) Slope C r a1 τ e τ s τ 2 t 13

14 Performance Characterization Delay and throughput characteristics were studied and comparisons where made with WFQ Simulations were performed using Extend, using discrete event simulation techniques Link speed of 1Mbps, mean packet lengths of 8000 bits (uniformly distributed service times) Poisson arrivals with exponential inter-arrival times were used All experiments were performed with three flows, with flow 1 always being the tagged flow (i.e. the flow whose load is varied) 14

15 Delay Results :HF 2 Q: Load < 1 The three flows had reservations of 0.2, 0.7 & 0.1 Mbps respectively The tagged flow load was varied and the incoming loads of flows 2 & 3 were 0.09 Mbps The tagged flow exceeds its reservation after a load of 0.2 Mbps Mean Delay (s) It is seen that HF Q behaves as 0.01 FIFO offering equal mean 0 waiting times for all flows Mean Delay for various incoming loads of tagged flow Flow 1 Flow 2 Flow 3 Incoming load for tagged flow (Mb/s) 15

16 Delay Results :WFQ: Load < 1 Same reservations and incoming loads were maintained for all flows It is observed that WFQ offers protection to the well-behaved flows (2 & 3) The delay of the tagged flow increases after it exceeds its reservation Flow 2 gets a lower delay compared to flow 3 since its reservation is higher Mean Delay (s) Flow 1 Flow 2 Flow 3 Mean Delay for various incoming loads of tagged flow Incoming load for tagged flow (Mb/s) 16

17 Delay Results :HF 2 Q: Load > 1 The tagged flow was increased so that the system load reaches 116% The delay of the tagged flow is not shown since it increases infinitely As is seen, the delay of the other two flows is low and bounded after a small transitional period As will be shown, HF 2 Q exhibits WFQ behavior after the transitional period Mean Delay (s) Mean Delay for various incoming loads of tagged flow Incoming load for tagged flow (Mb/s) Flow 2 Flow 3 17

18 Delay Results (Load > 1 cont.) Result for HF 2 Q (after the transitional region) Result for WFQ (after the transitional region) 0.03 Mean Delay for various incoming loads of tagged flow Mean Delay for various incoming loads of tagged flow Flow 2 Flow Mean Delay (s) Mean Delay (s) Flow 2 Flow Incoming load for tagged flow (Mb/s) Incoming load for tagged flow (Mb/s) 18

19 Throughput Experiment Throughput - fraction of link capacity used to carry packets Three flows with reservations of 0.5, 0.3 & 0.2 Mbps Flows 2 & 3 had incoming loads of 0.4 & 0.5 Mpbs respectively Flows 2 & 3 heavily exceed their reservation The system is overloaded after the incoming load of the tagged flow exceeds 0.1 Mbps The throughput behavior of FIFO, HF 2 Q and WFQ were studied 19

20 Throughput Results - FIFO All work conserving algorithms provide throughput equal to the incoming load when the total system load is less than unity When the system is overloaded FIFO offers throughput proportional to offered load WFQ and HF 2 Q on the other hand protects the wellbehaving 0.1 flows Normalized throughput Throughput vs Incoming load of tagged flow Incoming load of tagged flow (Mb/s) Flow 1 Flow 2 Flow 3 20

21 Throughput Results HF 2 Q & WFQ HF 2 Q WFQ Flow 1 Flow 2 Flow 3 Throughput vs Incoming load of tagged flow Throughput vs Incoming load of tagged flow Flow 1 Flow 2 Flow 3 Normalized throughput Normalized throughput Incoming load of tagged flow (Mb/s) Incoming load of tagged flow (Mb/s) 21

22 Conclusions Drawn HF 2 Q services packets in FIFO order when the total system load is less than one Indicated by equal mean waiting times Verified via simulations with packet sequence numbers Initial start-up period when ordering is not FIFO HF 2 Q services packets as WFQ does when heavily overloaded Transitional period when the load is barely over unity 22

23 DOCSIS Introduction In DOCSIS a Cable Modem Termination System (CMTS) controls the operations of many terminating Cable Modems (CMs). Upstream and downstream channels are separated using FDD. Each upstream channel is further divided into a stream of fixed-size time minislots (TDMA). DOCSIS MAC utilizes a request/grant mechanism to coordinate transmission between multiple CMs. 23

24 DOCSIS Operation 24

25 Request Mechanisms Contention - the CM may transmit a request in the contention period Piggybacking - is a request for additional bandwidth sent in a data transmission Unsolicited grants fixed size grants offered in a periodic basis Unicast request polls - unicast request opportunities are sent as a means of real-time polls regardless of network congestion 25

26 DOCSIS 1.1 QoS Classes Unsolicited Grant Service (UGS) Offers fixed unsolicited size grants on periodic basis Designed for fixed size data packet flows on fixed intervals Real Time Polling Service (rtps) Designed to support real-time flows that generate variable size packets on a periodic basis Offers periodic unicast polls that allow CM to specify the size of the desired grant 26

27 DOCSIS 1.1 QoS Classes (cont.) Non Real Time Polling Service (nrtps) Designed to support non-real time flows that require variable size grants on a regular basis Offers unicast polls on a regular basis and CMs are allowed to use contention opportunities and piggybacking Best Effort (BE) Service Provides efficient service to Best Effort Traffic Uses contention and piggybacking for requests Limited QoS Support 27

28 Performance Evaluation of DOCSIS 1.1 QoS Key performance attribute studied - Mean Access Delay Simulations were done in OPNET Comparison of BE and UGS performance Effect of DOCSIS 1.1 QoS features on performance Fragmentation - sending a portion of packet frame during a reserved slot time Concatenation - sending more than a frame during a transmission opportunity Piggybacking - a request carried in the next outgoing data frame Traffic Priority - CMTS uses Traffic Priority attribute for determining precedence in grant generation 28

29 Experiment 1: Comparison of Best Effort & UGS Delays UGS UGS flows are allowed to reserve certain portion of the bandwidth No transmission requests are needed; hence low, constant access delay BE request grant,request grant pattern Has to contend for sending requests May result in collision and thus increased delay due to retransmissions Piggybacking: requests are piggybacked to outgoing data and thus delay is reduced 29

30 Experiment 1: Comparison of Best Effort & UGS Delays (cont.) Load increased by adding BE stations Fragmentation & Concatenation were disabled Exponential packet lengths & inter-arrival times Backoff start = 7 30

31 Experiment 2: Effect of Fragmentation & Concatenation Load was increased by adding BE stations Exponential Packet Lengths and Exponential Inter-Arrival Times Fragmentation & Concatenation improve performance considerably Piggybacking was enabled;backoff start =7 31

32 Contention Resolution Algorithm (CRA) Overview CRA triggered by request collision Supported CRA: Truncated Binary Exponential Back-off Specified by an initial backoff window (Back-off start) and a maximum back-off (Back-off End) CM randomly selects a value within its back-off window - [0,2 Backoff ] Random number indicates the number of contention opportunities the CM must defer before transmitting requests On collision, increases the back-off window by a factor of 2 (less than back-off end) and repeats deferring process 32

33 Experiment 3: Effect of Backoff Start on Collision (piggybacking disabled) Packet lengths and interarrival times were made constant Since Piggybacking is disabled, collision probability increases with load As backoff start value increases, collision probability decreases 33

34 Experiment 3: Effect of Backoff Start on Collision (piggybacking enabled) Less Collision probability compared to the piggybacking disabled case since frequent contention is not necessary As load increases, collision probability decreases since contention load decreases 34

35 Experiment 3: Effect of Backoff Start on Delay Backoff 0 suffers maximum delay Backoff 4 produces a marked improvement in performance Backoff 7: reduced collision is offset by longer backoff delay 35

36 Experiment 4: Effect of Traffic Priorities on Delay CMTS uses Traffic Priority attribute for determining precedence in grant generation Priorities do not affect contention Traffic Priorities range from 0-7 applicable to BE, rtps & nrtps, 0 being the highest Hence high priority stations have lower access delays 36

37 Summary of Results Performance Evaluation of HF 2 Q HF 2 Q behaves as FIFO when the total system load is less than 1 Behaves as WFQ when the system is heavily overloaded Throughput behavior of HF 2 Q confirms these properties DOCSIS 1.1 UGS always receives bounded delay characteristics Fragmentation and Concatenation improve performance when system is overloaded Piggybacking improves delay characteristics CRA plays an important role in BE performance Traffic priority has significant effect on BE performance 37

38 Contributions & Future Work Performance characterization of a new queuing discipline -HF 2 Q was done Performance evaluation of DOCSIS QoS Scope for future work with regard to HF 2 Q Analytic explanation as to why HF 2 Q behaves as FIFO in an underloaded case Finding a closed form expression for the mean waiting time for WFQ Performance evaluation studies of two other traffic classes that DOCSIS 1.1 mentions namely rtps and nrtps yet to be done 38

39 Discussion Thank you! 39